I see that currently VecPow is used only in a small number of places including:
ksp/ksp/utils/lmvm/diagbrdn/diagbrdn.c: PetscCall(VecPow(ldb->U, ldb->beta - 1)); I am unsure if the usage here requires the special handling of negative numbers. I was wrong and it is not used in the semi-smooth code, that access the vector elements directly. If could be we can strip out all the special infinity cases completely. Barry > On Nov 14, 2024, at 9:56 AM, Stefano Zampini <stefano.zamp...@gmail.com> > wrote: > > That is a very old bug! Can you make an MR to just call PetscPowScalar in a > loop here > https://urldefense.us/v3/__https://gitlab.com/petsc/petsc/-/blob/main/src/vec/vec/utils/projection.c*L1022__;Iw!!G_uCfscf7eWS!b2Nfr7i01ZnyfVjweWqi87it8hcCDv0s6MtIsQhmSU8s4dq4jBPi-Ca87RRTwS20Srwyh9wQdhXcsbR6MVC8WSU$ > > <https://urldefense.us/v3/__https://gitlab.com/petsc/petsc/-/blob/main/src/vec/vec/utils/projection.c*L1022__;Iw!!G_uCfscf7eWS!bbSdSMnU5KpH03jHI7aV5j4WLGQ3yPvxWzR6Lwr14QLy_7EJ2MNT-qhL6J1x6z3vpF6M5GQk9lMQkMoLJ_JNSn1mqwqict4$> > ? > > Thanks > > Il giorno gio 14 nov 2024 alle ore 17:39 Peder Jørgensgaard Olesen via > petsc-users <petsc-users@mcs.anl.gov <mailto:petsc-users@mcs.anl.gov>> ha > scritto: >> Given a vector containing roots of unity, v[i] = exp(i*k*x[i]) I wanted to >> compute the vector u[i]=exp(i*n*k*x[i]), for some real number n. From the >> face of it this should be easily achieved with VecPow, as u[i] = v[i]^n. >> >> That didn't work as expected, though I got around it using VecGetArray() and >> a loop with PetscPowComplex(). The source designated in the docs >> (src/vec/vec/utils/projection.c) reveals that VecPow() maps v[i] to >> PETSC_INFINITY when the PetscRealPart(v[i]) < 0, unless the power is any of >> 0, ±0.5, ±1 or ±2. Even in the simple case of a purely real vector (with >> negative entries) raised to any other integer power, the results would not >> be what one might reasonably expect from the description of VecPow(). >> >> While I do have a solution suiting my need, I'm left wondering what might be >> the rationale for VecPow working the way it does. >> >> Best, >> Peder > > > > -- > Stefano
